COM for PAM4 Link Analysis–What you need to know
Geoff Zhang, Min Huang, Hongtao Zhang
SerDes Technology Group, Xilinx Inc. San Jose, CA 95124
Outline• COM in a nutshell• PAM4 in Brief
• Data Sampling Time in COM• DFE Tap Coefficients• Treatment of Jitter in COM• TX Output Level Separation Mismatch• CTLE in COM• COM Margin• DFE Error Propagation Impact• FEC Fundamentals • Summary and Future Work
2
COM in a Nutshell
• THRU and aggressor channels
• S-parameter representations
• Package models
• Single bit response (SBR)
• TX FFE: pre- & post- cursor taps
• RX CTLE transfer functions
• DFE and its tap coefficients
• Computation of FOM
• PDF’s of various impairments
• Computation of COM
• COM margin against DER0
3
PAM4 in Brief
4
• NRZ: 1 symbol = 1 bit; 2 levels and 1 eye
• PAM4: 1 symbol = 2 bits; 4 levels and 3 eyes
• Signal PSD for NRZ and PAM4
• PAM4 requires half the BW
• PAM4 symbol is twice wide
𝝈𝒙𝟐 =
2
𝑳∙
𝒌=1,3,…𝑳−1
𝒌
𝑳 − 1
2
= 1 for NRZ5/9 for PAM4
𝝈𝒙 = 0.7454
• PAR: Peak to Average Ratio
Data Sampling Time in COM
5
• Mueller-Muller baud-rate phase detector is emulated in COM
• After equalization 1st residual pre-cursor ≈ 1st residual post-cursor
• As long as b(1) does not limit (default <0.7), ℎ0 𝑡𝑠 − 𝑇𝑏 = 0
• This choice of ts reduces the impact of pre-cursor ISI. However
• It is usually at the expense of moving the sampling phase left of the SBR peak
• Increased 𝑏(1), due to both reduced ℎ0 𝑡𝑠 and increased ℎ0 𝑡𝑠 + 𝑇𝑏
• The consequence will be discussed in more details later
𝒉𝟎 𝒕𝒔
DFE Tap Coefficients
6
COM parameter for DFE specified in CEI-56G-LR-PAM4
• for 56G and 112G PAM4 most standards specify the raw BER between 1e-6 to 1e-4
• The impact of larger DFE tap values, together with multiple DFE taps, on real channel performance might be huge
Treatment of Jitter in COM
7
The variance of the amplitude
error due to timing jitter
• In COM jitter impact is converted to signal amplitude error before SNR is computed
The equalized SBR slope
for the THRU path
TX Output Level Separation Mismatch
8
• Level separation mismatch ratio, RLM
• RLM is typically specified > 0.95
• The distribution of 4 levels is unknown
𝐴𝑠 = 𝑅𝐿𝑀 ∙ ℎ 0 (𝑡𝑠)/(𝐿 − 1)
• 𝐴𝑠 is the available signal for NRZ
• For PAM4 it is scaled as below
CTLE in COM
9
• There are two stages of CLTE specified in COM for CEI-56G-LR-PAM4
𝐻𝐶𝑇𝐿𝐸 𝑓 = 𝑓𝑝2 ∙𝑗∙𝑓+𝑓𝑧∙10
𝐺𝐷𝐶20
𝑗∙𝑓+𝑓𝑝1 ∙ 𝑗∙𝑓+𝑓𝑝2
𝐻𝐶𝑇𝐿𝐸2 𝑓 = 𝑓𝑝2 ∙𝑗 ∙ 𝑓 + 𝑓𝐿𝐹 ∙ 10
𝐺𝐷𝐶220
𝑗 ∙ 𝑓 + 𝑓𝐿𝐹 ∙ 𝑗 ∙ 𝑓 + 𝑓𝑝2
When the peaking is excessive, the
design becomes practically very
challenging in one stage, thus more
parasitic poles should be added
HF stage LF stage
COM Margin
10
• FOM is firstly computed to decide the optimal equalization settings
𝑪𝑶𝑴 = 𝟐𝟎 ∙ 𝒍𝒐𝒈𝟏𝟎( 𝑨𝒔𝑨𝒏𝒊
)
• COM is computed based on the optimal settings
• The best FOM is not always equal to the best COM
• Ani is the total peak noise at DER0
𝑫𝑬𝑹𝟎 = 𝒌 ∙ 𝒆𝒓𝒇𝒄( 𝒉𝟎𝝈𝒕𝒐𝒕𝒂𝒍
𝟐)
• In COM k = 1/2 for NRZ and PAM4
• For PAM4 it should be k = 3/8
DFE Error Propagation Impact
11
• 1-tap DFE error propagation model
• 1-tap DFE error propagation probability • Average burst error length for 1-tap DFE
• Burst error definition
KP4 FEC Fundamentals
12
• FEC coding gain is the reduction in the required SNR that can be accommodated while still achieving the desired BER
• Under typical conditions, tests from system houses showed that KP4 FEC can achieve up to 8 dB in CG
• The exact CG value is a function of BER and error signature
• The KP4 FEC, RS(544, 514, T=15, M=10)
• In a codeword, 514 FEC symbols are encoded to form 544 FEC symbols
• Each FEC symbol contains M (=10) bits
• The FEC can correct up to T (=15) symbol errors within each codeword, regardless of number of bit errors
• This implies that
• At its most effective, KP4-FEC can correct as many as 150 bit errors in a codeword
• The other extreme, it can correct no more than 15 bit errors
DFE EP and FEC Correction Example
13
• 12-tap DFE setting for a case study
• Settings comply fully with CEI-56G-LR-PAM4
Base SERSER
with EP
SER ratio
after EP
Max burst
error length
Average burst
error length
Max KP4 FEC
symbol errors
1.0324e-6 3.3779e-6 3.2719 81 4.2757 17 (>15)
1.0050e-4 n/a n/a 2 2 6 (<16)
• Observations
• The SER increase after DFE is not alarmingly large; the raw BER is still better than specs
• The “average burst error length” is larger than the “SER ratio with EP” in the DFE case.
• The RS(544, 514) KP4 FEC fails completely even with limited sample size in the example
• Error signature is more relevant than BER itself in assessing FEC correction capability
• Without DFE EP the FEC can correct all the errors even though SER is two orders worse
Summary and Future Work
14
• COM is a great tool for qualifying a high speed serial link channels
• Focused on the COM for PAM4 signaling. We particularly analyzed • The potential impact from the sampling time and the DFE coefficient
limits; both could lead to severe error propagations
• KP4 FEC is discussed to provide a basic idea of error correction
• An example of DFE EP is provided to show that even DFE tap coefficients are well within the COM limits, the FEC can become dysfunctional
• Running COM to get the SNR margin does not guarantee the system to meet the desired performance
• Future work includes the following• DFE error propagation in a PAM4 link with multiple DFE taps
• Error signature and its impact on FEC coding gain analysis
• Precoding effect in terms of burst error removal
Thank you!
15